If it's a normal thing then it's hard to see how LN anodes can protect themselves from "influence attacks" (or whatever they're called) where a few nodes with a completely different and wrong blockchain connect as the only peers to another node, tainting its knowledge of what is the correct tip.
It's actually not even that easy to run a Lightning node without having Bitcoin Core running. Maybe it got easier, but last time I tried it was pretty bad.

Ended up just getting a disk of the required size for that node.
It's definitely essential for an LN node to know the status of the blockchain with full confidence. So either a full node runs on the same machine or the task is delegated to another machine which you trust (best control) and takes over the 'controlling' task - a watchtower, as Rath_ already said.
Do you have practical experience with autopilot mode plugins? Depending on implementation, it might drive centralization by going for the 'safe bet' (mostly going for channels towards large, well connected hubs).
I have only used LND's autopilot for a short moment back in 2019 and it always prioritized large nodes (especially ACINQ which always causes problems for me). I glanced over their GitHub and it looks like they have made quite a few improvements. I prefer to open my channels manually now.
That's good to know! It was kind of what I was expecting, since connecting to large hubs is the 'safe bet' strategy without too many what I call 'false positives' (opened channels that turn out useless).
Would be interesting to try if with all the code changes it got any better; I'm sure the developers of such plugins should be able to test their implementation through simulations (or better, the model based on which they write the logic code).